G
Gemma 4 31B
The most permissively licensed local model; 75.8% on science reasoning shows the ceiling for small models
Intelligence v5
— v4.3.2 19.0
Leaderboard rank
—
Input / output
$0.10/$0.40
Blended 3:1
$0.17
Output speed
200 tok/s
Value
—
Capability profile Normalised to the field maximum
Specifications
Vendor
Google
Released
2026-05
Weights
Open (Apache-2.0)
Context
262,144 tokens
Max output
Undisclosed
Parameters
31B dense
Input price
$0.10 / 1M
Output price
$0.40 / 1M
Cache hit
n/a
Output speed
200 tok/s
Intelligence
19.0（v4.3.2）
Benchmark scores
How it is measured →Bars are scaled to the field maximum for that benchmark.
| Benchmarks | Score | Vs. best in field | Gap |
|---|---|---|---|
| AA Intelligence Index v5 | — | — | not indexed |
| AA Intelligence Index v4.3.2 Intelligence index (legacy) | 19.0 | 33% | −38.6 pts |
| GPQA Diamond Graduate-level science reasoning | 75.8 | 79% | −20.6 pts |
| Terminal-Bench 4.0 | — | — | not indexed |
| LiveBench Design Document generation (DOCX/PPT/PDF) | 58.4 | 67% | −28.5 pts |
| Vals Index Vals composite index | 16.8 | 23% | −56.3 pts |
Assessment
Apache-2.0 in four sizes: 2B, 4B, 26B (A4B) and 31B. GPQA Diamond is 75.8% under the minimal-reasoning setting against 95%+ for the leaders — a 20-point gap that local deployment simply has to accept.
At the same price
The four models closest on blended price, side by side.
| Model | v5 | Blended | tok/s | Score gap |
|---|---|---|---|---|
| Claude Haiku 5.5Anthropic | 36.2 | $0.20 | 235 | — |
| MiniMax M3.1 FlashMiniMax | 29.0 | $0.26 | 118 | — |
| DeepSeek V4.1 Flash Open weightsDeepSeek | 32.5 | $0.53 | 148 | — |
| MiMo-V2.6 Pro Open weightsXiaomi | 30.0 | $0.61 | 81 | — |
Compare side by side
Line it up against other models on price, speed and every benchmark.
The comparison tool lines up to 4 models side by side with switchable normalisation — built for purchasing decisions.
Open the comparison tool → Back to leaderboard